NCLT dataset split

[alexmelekhin]

Can you provide code for making train/test splits for NCLT dataset or at least prepared train/test pickles? Without those file it is impossible to make a fair comparison of different methods.

[sijieaaa]

+1

[HaowenLai]

Thanks for the interest in our work. For the RobotCar dataset, we specified four 150m x 150m regions for testing and the rest for training. For the NCLT dataset, we specified four 100m x 100m regions for testing.

These four testing regions can be determined by four lower left starting points $p(x, y)$. Then the upper right points are ether $p'(x+150, y+150)$ or $p'(x+100, y+100)$, depending on which dataset you use.

The four starting points are:

• RobotCar dataset:

    x_width = 150
    y_width = 150
    p1 = (5735712.768124, 620084.402381) # (northing, easting)
    p2 = (5735611.299219, 620540.270327)
    p3 = (5735237.358209, 620543.094379)
    p4 = (5734749.303802, 619932.693364)

• NCLT dataset:

    x_width = 100
    y_width = 100
    p1 = (-260.0, -680.0) # (northing, easting)
    p2 = (-280.0, -420.0)
    p3 = (20.0, -550.0)
    p4 = (-100.0, -300.0)

[sijieaaa]

Thanks for your reply! But it is still not clear to reproduce the results. Could you provide "train_pairs_10.pickle" and "test_items_10.pickle" that you used in your code? https://github.com/MetaSLAM/AdaFusion/blob/b9f9b7abb1f53f98ea7333fb74383a0ffa6330c9/experiments/aavislidar3_pairwise/config.yaml#L22

Thanks!